Chinese AI models are approaching the intelligence inflection point for global proliferation

Overview

This report is Goldman Sachs' in-depth thematic research on China's AI model industry. Its core judgment is that Chinese open-source and open-weight models are approaching a critical point where their intelligence performance nears that of leading global proprietary models, while expanding adoption among domestic enterprises and global SMEs through lower inference costs, smaller parameter sizes, and architectural innovations such as MoE. The report focuses on how Chinese models achieve competitive performance at low cost, why they adopt the open-source/open-weight route, the addressable market and monetization path, and which companies are more likely to become long-term winners.

Core views

The report believes that the Chinese AI model market is forming a two-tier structure: high-performance models retain pricing power through intelligence level and speed to market, while low-cost agent models unlock price-sensitive demand at US$0.06-0.2/1M blended tokens. Real usage growth of Chinese models in coding and agent scenarios will form a data flywheel and reduce reliance on distillation. Long-term winners should have the largest ARR scale, gross margin advantage, and financial strength, with Zhipu and DeepSeek strongest in text foundation models and ByteDance leading in multimodal.

Analysis framework

The report combines an industry framework with company comparisons, analyzing the long-term competitive positioning of Chinese AI model companies around token scale, pricing, inference cost, model architecture, compute availability, cash reserves, valuation multiples, and market share, while separately incorporating domestic and overseas demand, API and subscription revenue pools, and coding/agent/multimodal applications into the discussion.

Methodology notes

• Competitive positioningCompetitive Positioning framework

  Pricing power, cost advantage, financial strength

  The report evaluates AI model companies across three dimensions: pricing power is measured by model launch timing, Arena real-usage scores, and price level; cost advantage is measured by token scale, throughput, cache hit rate, parameter size, activation ratio, and estimated inference gross margin; financial strength is measured by cash, net cash as a share of assets, and valuation multiples.

• Market sizeToken and revenue share forecast

  Token growth and revenue pool forecast

  The report estimates domestic and overseas revenue opportunities for Chinese AI model companies using token consumption, API pricing, subscription revenue, and market share, and extrapolates for uncovered companies such as DeepSeek, ByteDance, and Zhipu using industry and covered-company data.

• Business modelOpen-source/open-weight monetization

  Open-source proliferation and commercial licensing

  The report believes open source helps training, deployment flexibility, and community adoption, but that future high-performance models may shift more toward open weight plus Community License, monetizing through commercial-use revenue sharing or take rate.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Zhipu / Knowledge Atlas

  One of the leaders in text foundation models

  Strengths

  Strong coding and agent model capabilities; models such as GLM5.2 support pricing improvement, and the report lists it as one of the strongest-positioned companies in foundation models.

  Weaknesses

  As an independent AI company, it still needs to continue proving commercialization scale, financial strength, and access to high-end compute.

  Comparison

  Listed together with DeepSeek as the strongest in text foundation models; compared with mega-caps, its capital and ecosystem resources may be weaker.

  Risks

  Overseas model access restrictions, catch-up by competitors with larger-parameter models, pricing pressure, and regulatory changes.

• DeepSeek

  Representative of low-cost, high-efficiency foundation models

  Strengths

  Outstanding cost efficiency; DSpark improves the online service speed of DeepSeek-V4 Flash / Pro, supporting scalable token usage.

  Weaknesses

  Some indicators such as valuation multiples are not applicable or have limited disclosure, and commercialization and market access remain uncertain.

  Comparison

  Along with Zhipu, it is among the strongest-positioned foundation models; it is especially strong in cost efficiency.

  Risks

  Overseas access restrictions, anti-distillation policies, access to high-end training compute, and low-price competition.

• ByteDance

  Leader in multimodal models

  Strengths

  The report believes ByteDance leads in multi-modal capabilities, and Seedance ARR run-rate and gross margin performance are healthy.

  Weaknesses

  Relevant operating metrics are extrapolated by the report from industry and covered-company data, with limited disclosure transparency.

  Comparison

  Compared with text foundation models, ByteDance stands out more in multimodal and video generation.

  Risks

  Overseas regulation, data security reviews, local compute requirements, and market access restrictions.

• MiniMax Group (0100.HK)

  Beneficiary of multimodal and low-cost models

  Strengths

  M3 is in the report's preferred ARR maximization quadrant, with outstanding cost efficiency and a high overseas revenue mix; the report reiterates Buy.

  Weaknesses

  Its overall competitive positioning still depends on further improvements in pricing power and financial strength.

  Comparison

  Valued at 13X P/2026E year-end ARR; the report says it trades at a discount relative to Chinese and global ARR peers at a similar stage.

  Risks

  Low-end API pricing pressure, model update timing, H3 video generation performance, and funding capacity.

• Alibaba Qwen

  Mega-cap beneficiary of full-stack AI and cloud

  Strengths

  Synergies between full-stack AI and cloud business; high-performance Qwen models have pricing power of around US$1/1M blended tokens.

  Weaknesses

  Cloud capex conversion efficiency, model competition, and price wars will affect the pace of returns.

  Comparison

  Compared with independent model companies, Alibaba has cloud and cash flow support, but it must continue to lead in model performance and ecosystem adoption.

  Risks

  Capex returns, cloud competition, regulation, and high-end compute restrictions.

• Tencent Hunyuan

  Mega-cap AI model competitor

  Strengths

  The report says it is a strong competitor after its AI strategy adjustment, and the Weixin AI agent has potential.

  Weaknesses

  Model ranking, commercialization speed, and ecosystem closure still need to be validated.

  Comparison

  Like Alibaba, it is a mega-cap player backed by funding support, but its application entry points and social traffic differ significantly.

  Risks

  Lagging model iteration, weaker-than-expected productization, and price competition.

Key data

• China AI model token growth25X by 2030EThe report expects tokens generated by Chinese AI model companies to increase 25-fold by 2030E versus current levels.
• Overseas token share5% of global tokens by 2030EThe report expects tokens generated overseas by Chinese AI model companies to reach 5% of global tokens by 2030E.
• China AI API + subscription revenue poolUS$125bn in 2030EThe report estimates China's AI API and subscription revenue pool will reach US$125bn by 2030E.
• High-performance Chinese model pricingaround US$1 per blended 1M tokensHigh-performance models such as Zhipu GLM5.2 and Alibaba Qwen3.7 Max are priced at around 5X the low-end Chinese model level.
• Low-end agent model pricingUS$0.06-0.2 per blended 1M tokensThe low-price range helps models reach price-sensitive global SMEs and individual company demand.
• Chinese model parameter scale200bn to 1.6T parametersThe report says Chinese model parameter scales are about 2-10% of leading SOTA models and improve efficiency through MoE, Sparse Attention, and related methods.
• LongCat 2.01.6 trillion-parameter open-source MoE modelMeituan LongCat 2.0 is seen as an important milestone in China's domestic AI infrastructure, reportedly trained and deployed on a 50,000-card domestic compute cluster.
• ByteDance Seedance ARRUS$2bn+ ARR run-rate; 70% gross marginCiting LatePost and 36Kr, the report says ByteDance Seedance's latest ARR run-rate exceeds US$2bn, with gross margin around 70%.

Impact & implications

The investment implication is that the value focus of China's AI model industry may shift from pure token-volume expansion toward ROI, task cost, real agent usage, and enterprise automation outcomes. High-performance coding models, multimodal video generation, and low-cost agent models each correspond to different ARR maximization paths; companies with advantages in compute, cash, and inference efficiency are more likely to retain long-term value after the price war.

Risks

• Access restrictions in overseas markets for leading Chinese AI models may tighten.
• Data security, deployment, and compliance policies in Western markets toward Chinese models may restrict proliferation.
• Restricted access to high-end training compute and overseas rented compute could affect model iteration.
• Price wars in low-end agent model APIs may compress gross margins and prolong the profitability inflection point.
• Anti-distillation, entity list, and market access policies may alter the industry's growth trajectory.
• SLMs and new AI architectures may weaken the long-term value of the current large-model route.
• Multimodal demand is strong but constrained by compute capacity, and supply-demand mismatches may affect service stability.

What to watch

• Whether Chinese model companies launch new 2-5 trillion parameter-class models in 2H 2026.
• How subsequent updates to Zhipu GLM, DeepSeek V4, and MiniMax M3/H3 improve coding, agent, and video generation capabilities.
• Whether API pricing shifts from token-maxxing toward ROI, task cost, and Daily Active Agents as key measures.
• Whether Chinese domestic ASICs and local compute clusters can support more frontier training.
• Whether overseas cloud vendors host Chinese open-weight models, and whether regulation tightens.
• Whether China's AI API + subscription revenue pool and overseas token share materialize in line with the report's forecasts.
• ARR and gross margin trends for video generation models such as ByteDance Seedance, Kuaishou Kling, and MiniMax Hailuo/H3.